Supercomputers | Research & Encyclopedia Articles

Supercomputers

Supercomputing is a concept that has evolved rapidly. Computers developed under that description have existed only since the early 1970s. For most of the time since, supercomputers have been widely imagined to be expensive, arcane devices, employed in only the most esoteric functions of theoretical science and government research.

Since the mid-'90s, however, supercomputers have become more efficient, less expensive, and available for applications in all walks of life. Supercomputers are currently relied upon for computer-generated animation, for advanced design of automobiles and other complicated machinery, and in the more complex areas of medicine such as microbiology. These are in addition to the established roles of supercomputing in creating weather models, making population-growth projections and in digitally simulating nuclear detonations.

So what is a supercomputer? No universally accepted definition exists, other than that of a vague class of exceptionally fast computers. Parallel processing, that is, the practice of directing multiple processors toward the completion of a common task, is usually identified closely with supercomputing. It is true that most supercomputers make use of parallel processing, or its cousin, vector processing. But since supercomputing predates parallel processing, and since it is likely that coming generations of supercomputers might evolve beyond it, a working definition of supercomputing must embrace a wider scope. Supercomputers, then, are simply whichever computers are the most powerful of a given generation.

Although a widening field of manufacturers are engaged in supercomputer production--among these are IBM, Fujitsu Ltd., NEC Corporation, Intel, Hitachi Ltd., and Thinking Machines Corporation--a single name is most often identified with both the birth of supercomputing and with the state of the art. Seymour Cray (1925-96), backed by the handful of companies he founded, has consistently developed and delivered some of the most advanced supercomputing concepts and practices in the world. From the installation of the first Cray-1 at Los Alamos National Laboratory in 1976, to the creation of the Cray-4 (the first supercomputer which was smaller than a human brain), Cray endeavored always to advance the science of supercomputing with only the most cursory attention paid to commercial success.

The key to Cray's designs have always been efficiency. Cray has been adept at tapping the most advanced technology of the era in the name of better and faster computers. One of his earliest systems, the CDC1604 created for Control Data Corporation in the late 1950s, was the first transistorized scientific computer--developed at a time when all IBM computers were still utilizing punch-card systems. Later, in 1972, Cray Research was the first to utilize the integrated circuit in computer design. Cray systems tend to be designed with physical efficiency in mind as well. The "horseshoe" shape of several of the Cray systems was intended to bring all components closer together, and allowed for all elements of wiring to be no longer than four feet in length. Cray has been innovative in peripheral systems as well, such as cooling. Early Cray designs incorporated their own Freon reservoirs, which might have been only partially successful when one considers that employees of Cray Research in Minnesota were known to run a supercomputer or two simply to heat the office. Later designs, such as the Cray-4, utilized as a coolant the same chemical used in medical applications as artificial blood. Finally, Cray has even innovated the technology of chip manufacturing. Since the early '90s, Cray has eschewed the use of silicon in favor of gallium arsinide (GaAs). While gallium arsinide has proven to be a more efficient platform for circuitry, its utility is so complicated that Cray Research was forced to build their own gallium arsinide foundry.

Historically, the world's most powerful computers were gauged by their ability to complete instruction sets--measured in MIPS (millions of instructions per second). As software became more and more distinct from the hardware, however, it became clear that the MIPS measurement could too easily be skewed by vagaries in the software code. A new benchmark, based upon pure computing power, had to be formed.

The current standard is the FLOP, or floating-point operation. Today's supercomputers are classified by their capabilities in floating-point operations per second in the millions (megaflop), billions (gigaflop) and most recently, trillions (teraflop). The teraflop barrier was first broken in the mid-'90s, although there is some contention as to which supercomputer was the first to sustain operations in the teraflop range. It's interesting to note that a truly unconventional design may hold pride of place as the world's largest supercomputer, operating in the seven-teraflop range. SETI (Search for Extraterrestrial Intelligence), upon having their funding cut by the U.S. Congress in the early '90s, issued a worldwide appeal for computer users to donate their systems' downtime for the analysis of billions of intercepted stellar radio waves. More than one million processors in 223 countries make up the components of this global "supercomputer."

Like all computer manufacturers, the world's supercomputer providers are engaged in an undeclared yet ongoing competition for faster and better machines. Most have 100 teraflop systems on their drawing boards that could be ready for distribution in a decade or less. And before he died in an automobile accident in 1996, Seymour Cray reported a government initiative to conquer the petaflop (10¹⁵ FLOPS) realm. As unimaginable as that number might be, the history of supercomputing--with all its innovations and its penchant for defying the odds--instructs us to discount no such possibilities.